71 Commits

Author	SHA1	Message	Date
hiyouga	92dab8a90b	simplify readme	2024-04-02 20:07:43 +08:00
hiyouga	b267aeb53f	add moe aux loss control #3085	2024-04-02 14:26:31 +08:00
hiyouga	9ddbe2866a	fix #3022	2024-04-02 13:58:39 +08:00
hiyouga	4a6ca621c0	fix #3083	2024-04-01 22:53:52 +08:00
hiyouga	aee634cd20	fix #3077	2024-04-01 21:35:18 +08:00
hiyouga	eb259cc573	support infer 4bit model on GPUs #3023	2024-04-01 17:34:04 +08:00
hiyouga	27776c3474	tiny fix	2024-03-31 00:10:29 +08:00
marko1616	d9a5134617	fix blank line contains whitespace	2024-03-30 23:46:55 +08:00
marko1616	eb178eaff3	Fix Llama model save for full param train	2024-03-30 23:45:04 +08:00
hiyouga	511f675402	fix #2961	2024-03-26 17:26:14 +08:00
hiyouga	a1c8c98c5f	fix #2941	2024-03-24 00:28:44 +08:00
hiyouga	8408225162	support fsdp + qlora	2024-03-21 00:36:06 +08:00
hiyouga	7b8f502901	fix #2346	2024-03-20 17:56:33 +08:00
hiyouga	85c376fc1e	fix patcher	2024-03-15 19:18:42 +08:00
S3Studio	e75407febd	Use official Nvidia base image ... Note that the flash-attn library is installed in this image and the qwen model will use it automatically. However, if the the host machine's GPU is not compatible with the library, an exception will be raised during the training process as follows: FlashAttention only supports Ampere GPUs or newer. So if the --flash_attn flag is not set, an additional patch for the qwen model's config is necessary to set the default value of use_flash_attn from "auto" to False.	2024-03-15 08:59:13 +08:00
hiyouga	3b4a59bfb1	fix export	2024-03-14 18:17:01 +08:00
hiyouga	72367307df	improve lora+ impl.	2024-03-13 23:32:51 +08:00
hiyouga	96ce76cd27	fix kv cache	2024-03-13 01:21:50 +08:00
hiyouga	8d8956bad5	fix #2802	2024-03-12 17:08:34 +08:00
hiyouga	bdb496644c	allow non-packing pretraining	2024-03-09 22:21:46 +08:00
hiyouga	e8dd38b7fd	fix #2756 , patch #2746	2024-03-09 02:01:26 +08:00
hoshi-hiyouga	516d0ddc66	Merge pull request #2746 from stephen-nju/main ... fix deepspeed ppo RuntimeError	2024-03-09 01:37:00 +08:00
hiyouga	10be2f0ecc	fix aqlm version	2024-03-09 00:09:09 +08:00
stephen_zhu	aa71571b77	update	2024-03-08 12:47:44 +08:00
stephen	cdb7f82869	fix ppo runtime error	2024-03-08 11:48:26 +08:00
hiyouga	f74f804a71	fix #2735	2024-03-07 16:15:53 +08:00
hiyouga	3e84f430b1	export use balanced gpu	2024-03-06 16:33:14 +08:00
hiyouga	d3d3dac707	optimize aqlm training	2024-03-05 18:35:41 +08:00
hiyouga	ddf352f861	fix dora inference	2024-03-05 11:51:41 +08:00
hiyouga	cda2ff8727	fix export on cpu device	2024-03-04 17:35:09 +08:00
hiyouga	4e5fae2fac	fix #2649	2024-03-01 13:02:41 +08:00
hiyouga	c0be617195	fix #2642	2024-02-29 18:32:54 +08:00
hiyouga	4a871e80e2	tiny fix	2024-02-29 17:28:50 +08:00
hiyouga	fa5ab21ebc	release v0.5.3	2024-02-29 00:34:19 +08:00
hiyouga	cfefacaa37	support DoRA, AWQ, AQLM #2512	2024-02-28 19:53:28 +08:00
hiyouga	7924ffc55d	support llama pro #2338 , add rslora	2024-02-15 02:27:36 +08:00
hiyouga	0ae9a16b9d	update gc kwargs	2024-02-07 00:38:24 +08:00
hiyouga	ebf31b62eb	fix #2438	2024-02-06 15:23:08 +08:00
hiyouga	4ecadc3512	fix #2376	2024-02-03 23:14:31 +08:00
hiyouga	521ad76552	fix autoset attn impl, update data readme	2024-01-31 11:58:07 +08:00
hiyouga	2bc30763e9	fix #2320	2024-01-24 16:19:18 +08:00
ldwang	c284665425	Add patch_mixtral_replace_moe_impl for full training Mitral using DeepSpeed Zero3. ... Signed-off-by: ldwang <ftgreat@gmail.com>	2024-01-24 15:25:31 +08:00
ldwang	18923b1402	Add patch_mixtral_replace_moe_impl for full training Mitral using DeepSpeed Zero3. ... Signed-off-by: ldwang <ftgreat@gmail.com>	2024-01-24 14:43:16 +08:00
hiyouga	e4ba1deedf	add hint	2024-01-22 23:32:01 +08:00
hoshi-hiyouga	bdc9eff635	Update patcher.py	2024-01-22 23:27:39 +08:00
A-Cepheus	b06a31e76a	🐞 fix: typo	2024-01-22 16:04:39 +08:00
A-Cepheus	319a72b48d	🐞 fix: typo, move MoE fix to patcher	2024-01-22 16:01:58 +08:00
hiyouga	e0a717aa3a	fix #2268	2024-01-21 14:11:38 +08:00
hiyouga	638234ceee	format style	2024-01-20 20:15:56 +08:00
hiyouga	38af076a75	support longlora for main branch	2024-01-20 19:25:22 +08:00